From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752767AbdK3Nno convert rfc822-to-8bit (ORCPT ); Thu, 30 Nov 2017 08:43:44 -0500 Received: from mx1.redhat.com ([209.132.183.28]:41412 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752347AbdK3Nnm (ORCPT ); Thu, 30 Nov 2017 08:43:42 -0500 Subject: Re: [PATCH] list_lru: Prefetch neighboring list entries before acquiring lock To: Minchan Kim Cc: Andrew Morton , Vladimir Davydov , Johannes Weiner , Dave Chinner , linux-kernel@vger.kernel.org, linux-mm@kvack.org References: <1511965054-6328-1-git-send-email-longman@redhat.com> <20171130005301.GA2679@bbox> From: Waiman Long Organization: Red Hat Message-ID: <414f9020-aba5-eef1-b689-36307dbdcfed@redhat.com> Date: Thu, 30 Nov 2017 08:43:41 -0500 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.2.0 MIME-Version: 1.0 In-Reply-To: <20171130005301.GA2679@bbox> Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 8BIT Content-Language: en-US X-Greylist: Sender IP whitelisted, not delayed by milter-greylist-4.5.16 (mx1.redhat.com [10.5.110.30]); Thu, 30 Nov 2017 13:43:42 +0000 (UTC) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 11/29/2017 07:53 PM, Minchan Kim wrote: > Hello, > > On Wed, Nov 29, 2017 at 09:17:34AM -0500, Waiman Long wrote: >> The list_lru_del() function removes the given item from the LRU list. >> The operation looks simple, but it involves writing into the cachelines >> of the two neighboring list entries in order to get the deletion done. >> That can take a while if the cachelines aren't there yet, thus >> prolonging the lock hold time. >> >> To reduce the lock hold time, the cachelines of the two neighboring >> list entries are now prefetched before acquiring the list_lru_node's >> lock. >> >> Using a multi-threaded test program that created a large number >> of dentries and then killed them, the execution time was reduced >> from 38.5s to 36.6s after applying the patch on a 2-socket 36-core >> 72-thread x86-64 system. >> >> Signed-off-by: Waiman Long >> --- >> mm/list_lru.c | 10 +++++++++- >> 1 file changed, 9 insertions(+), 1 deletion(-) >> >> diff --git a/mm/list_lru.c b/mm/list_lru.c >> index f141f0c..65aae44 100644 >> --- a/mm/list_lru.c >> +++ b/mm/list_lru.c >> @@ -132,8 +132,16 @@ bool list_lru_del(struct list_lru *lru, struct list_head *item) >> struct list_lru_node *nlru = &lru->node[nid]; >> struct list_lru_one *l; >> >> + /* >> + * Prefetch the neighboring list entries to reduce lock hold time. >> + */ >> + if (unlikely(list_empty(item))) >> + return false; >> + prefetchw(item->prev); >> + prefetchw(item->next); >> + > A question: > > A few month ago, I had a chance to measure prefetch effect with my testing > workload. For the clarification, it's not list_lru_del but list traverse > stuff so it might be similar. > > With my experiment at that time, it was really hard to find best place to > add prefetchw. Sometimes, it was too eariler or late so the effect was > not good, even worse on some cases. > > Also, the performance was different with each machine although my testing > machines was just two. ;-) > > So my question is what's a rule of thumb to add prefetch command? > Like your code, putting prefetch right before touching? > > I'm really wonder what's the rule to make every arch/machines happy > with prefetch. I add the prefetchw() before spin_lock() because the latency of the lockinig operation can be highly variable. There will have high latency when the lock is contended. With the prefetch, lock hold time will be reduced. In turn, it helps to reduce the amount of lock contention as well. If there is no lock contention, the prefetch won't help. Cheers, Longman